PRemiSE: Bridging the "Word of Mouth" and Matrix Factorization for News Recommendation

TL;DR

The news cycle moves at a breakneck pace, rendering traditional collaborative filtering (CF) nearly useless for brand-new stories or anonymous readers. PRemiSE (Personalized news Recommendation via implicit Social Experts) solves this by identifying "virtual experts" within a reading community. By blending content analysis with the "word of mouth" effect into a Probabilistic Matrix Factorization (PMF) framework, it achieves superior accuracy and solves the notorious cold-start problem.

The Core Challenge: The Ephemeral Nature of News

Most recommendation systems rely on long-term user-item interactions. In news, however:

• Data Sparsity: Users read only a tiny fraction of daily news.
• Double Cold-Start: New articles arrive every hour (item cold-start), and many readers are unregistered or first-time visitors (user cold-start).
• Lack of Explicit Social Data: Unlike Twitter or Facebook, news portals don't usually have "friend" or "follow" graphs to rely on.

The authors' insight is simple: even without an explicit social network, a virtual social network exists. If a small group of "experts" reads a niche technical story and it eventually spreads, their early adoption serves as a signal for others.

Methodology: The PRemiSE Framework

PRemiSE doesn't just look at what you liked; it balances your personal profile with the influence of community experts.

1. Hybrid Factorization

In standard PMF, a rating is the product of user factor $U$ and item factor $V$. PRemiSE modifies this by defining the user's preference as a weighted combination:

• $w_1$: Personal preference (dominant for experienced users).
• $w_2$: "Word of mouth" from experts (dominant for new users).

2. Semantic Integration

To handle new items, the model maps news content (TF-IDF vectors) directly into the latent factor space. This allows the system to estimate an item's factor $V_j$ based on its words before anyone has even clicked on it.

Figure 1: Comparison between basic PMF (a) and the PRemiSE framework (b) which includes social expert influence and content semantics.

Experiments and Cold-Start Mastery

The authors tested the model against standard CF, basic Matrix Factorization, and LDA-based filtering.

Key Performance Indicators:

• Accuracy: PRemiSE consistently yielded lower RMSE (Root Mean Square Error) across multiple datasets (Stories and Entities).
• Cold-Start Resilience: While baselines like CF and MF struggle to make meaningful predictions for new users, PRemiSE uses the "Global Expert" opinion as a fallback, maintaining high recommendation quality.

Figure 3: Breakdown of results for Existing-User/Existing-Item (oop) vs. Cold-Start scenarios (onp, nop, nnp). PRemiSE shows a clear advantage in new item/user prediction.

Critical Insight: Why Experts?

The "Wisdom of the Few" is often more effective than the "Wisdom of the Crowd" in specialized domains. By automatically detecting experts—those who have a high influence on the information diffusion within the community—PRemiSE creates a robust "anchor" for recommendations when individual user data is missing.

Conclusion

PRemiSE demonstrates that social influence is a powerful latent feature even when no explicit "Follow" button exists. By integrating content, collaboration, and implicit social structures, the framework provides a blueprint for building resilient recommenders in high-churn environments like digital journalism.

Future Outlook: Integrating these "implicit experts" with modern transformer-based embeddings could potentially push the boundaries of zero-shot personalization even further.